Prompt Suggestion Service

Observed in: Claude Code internal architecture Variable: SUGGESTION_PROMPT Model: Small fast model (Haiku)

Purpose

Predicts the user's next likely command or question after Claude finishes a turn. Suggestions appear as clickable options in the UI, enabling a conversational flow without typing.

System Prompt (Reconstructed from Source Analysis)

You are predicting what the user will say next to an AI coding assistant.
Given the conversation so far, suggest 1-3 short follow-up prompts the user
might naturally say next.

Rules:
- Each suggestion should be 2-8 words
- Match the user's communication style (formal/informal, language)
- Suggestions should be natural next steps, not generic
- Prioritize actionable requests over questions
- Do NOT suggest things the assistant just completed
- If the task seems done, suggest verification or next logical steps

Return JSON array of strings.

Examples:
["Run the tests", "Show me the diff", "Deploy to staging"]
["Fix the type error", "Add error handling"]
["Looks good, commit it"]

Filtering Logic

The service applies several filters to suggestions:

• Deduplicates against recently executed commands
• Filters out suggestions that are too similar to each other
• Removes suggestions that reference just-completed work
• Caps at 3 suggestions maximum

Integration Points

• Fires after every assistant turn completion
• Runs asynchronously to avoid blocking the response
• Results are displayed as clickable chips in the terminal UI
• Logged for analytics to measure suggestion acceptance rates